Sensitivity analysis and uncertainty propagation of the time to onset of natural circulation in air ingress accidents

This study investigates the time to onset of natural circulation (ONC) during a depressurized loss of forced cooling (DLOFC) event in a high-temperature gas reactor (HTGR). Using a fully automated Ansys Fluent simulation framework with PyFluent scripting, 500 CFD cases were generated with perturbed thermal and material properties. Surrogate models (random forests and neural networks) were trained to predict ONC time and post-ONC temperature, enabling global sensitivity analysis (SA) via Morris screening, Sobol indices, Fourier Amplitude Sensitivity Test, and regional sensitivity analysis. Monte Carlo-based uncertainty quantification was performed using the trained surrogates. Results showed that the heated section temperature was the dominant factor influencing ONC timing, with negligible contributions from heat transfer coefficient (HTC) and other thermophysical properties. In contrast, post-ONC temperature was influenced by both initial temperature and HTC. Sensitivity analysis revealed signs of nonlinear behavior and potential interactions between these parameters. The neural network achieved a test R^2 of 0.986 and MAE of 64 s for ONC timing, and an R^2 of 0.993 and MAE of 10 K for post-ONC temperature. While the random forest performed slightly worse, it still achieved a test R^2 of 0.985 and MAE of 64 s for ONC timing, and an R^2 of 0.964 with MAE of 24 K for post-ONC temperature. Using these surrogate models, the uncertainty propagation results verified the influence of the primary input parameters identified by sensitivity analysis on ONC timing and post-ONC temperature.